Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Contact-Wise SEEG Decoding Reveals Sustained and Spatially Distributed Phoneme Representations During Speech Perception
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Aidan Truel1,2, Mamady Nabe3, Laurent Spinelli4, Pierre Mégevand3,4, Timothée Proix2; 1University of Zurich, Zurich, Switzerland, 2ETH Zurich, Zurich, Switzerland, 3University of Geneva, Geneva, Switzerland, 4Geneva University Hospitals, Geneva, Switzerland
Phonemes are the smallest contrastive speech units that change a word’s meaning. Perceiving distinct phonemes and maintaining them in memory is crucial for speech comprehension. Previous work has shown that multiple phonemes are represented in parallel, long past the dissipation of the acoustic signal (Gwilliams et al., 2022) and that specific brain areas respond selectively to distinct phonetic properties (Mesgarani et al., 2014). However, the timing and duration of their representations may vary across brain regions depending on their hierarchical function in language processing. To investigate this, we introduce a single-contact intracranial decoding approach that quantifies the timing and duration of phonetic representations while preserving spatial specificity. We analyzed long-term stereo-EEG (SEEG) recordings from two French-speaking patients under clinical monitoring prior to epilepsy surgery. Broadband high-frequency activity (BHA) was extracted from all electrode contacts (>150 per patient) and aligned to natural speech recorded over several days. Phoneme-level annotations were grouped into four categories (consonant manner and place of articulation, vowel height and backness), and linear decoders were trained to discriminate each pair of phonetic contrasts within a category. For each contact, decoders were evaluated at multiple temporal windows (50-500ms length) at various times relative to phoneme onset (-500 to +500ms in 25ms steps). Contact-contrast pairs were retained for analysis if their decoding accuracy exceeded a fixed patient-specific threshold for a fixed duration. We developed the following method to quantify the duration and timing of neural representations. For each retained contact-contrast pair, decoding time courses for all window lengths were aligned separately to window start and end times, then summed. We used the peak of the start-aligned and end-aligned curves as estimates of the onset and offset of representation, respectively. The duration was quantified as the time lag that maximized the Pearson correlation between the start-aligned and end-aligned curves. In both patients, phonetic contrasts were decodable from many distinct regions. Vowel backness showed the most widespread representation, surviving thresholding in the most contacts. Place-of-articulation contrasts were also represented in more contacts than manner-of-articulation contrasts. Onset and offset of representation revealed a spatial progression unique to each patient. The duration ranged from 340-500ms, with region-specific and patient-specific variations. This slightly exceeds prior duration estimates from MEG studies (300ms in Gwilliams et al., 2022) and far exceeds typical phoneme duration (70-120ms), necessitating the simultaneous representation of multiple phonemes. Notably, several regions, including superior temporal gyrus (STG) and rostral middle frontal gyrus (rMFG), shared highly similar timing between patients. We introduced a new method to quantify the temporal structure of neural representations from single-contact decoding results, demonstrating that phonetic information is sustained in neural activity far longer than phoneme duration and is represented with variable timing across brain regions.
Topic Areas: Speech Perception, Computational Approaches